Morphological Processing in the Nabu System

(1)

M O R P H O L O G I C A L P R O C E S S I N G I N T H E N A B U S Y S T E M

Jonathan Slocum Microelectronics and C o m p u t e r Technology Corporation (MCC) 3500 West Balcones Center Drive

Austin, Texas 78759 A B S T R A C T

The N a b u morphological processor is designed to perform a number of different functions, of which five have so far been identified: analysis, guessing (about u n k n o w n words), synthesis, defaulting (proposing the most likely inflectional paradigm for a new base form), and coding (producing all possible inflectional paradigm variants for a new base form). C o m - plete or very substantial analyzers have been produced for a number of diverse languages; other functions have been implemented as well. This paper discusses our design philosophy, as well as our technique and its implementation.

I N T R O D U C T I O N

Nabu is a multilingual Natural Language Processing system under development in the H u m a n Interface L a b o r a t o r y at MCC, for shareholder companies. Its morphological component is designed to perform a number of different functions. This has been used to produce a complete analyzer for Arabic; very substantial analyzers for English, French, German, and Spanish; and small collections of rules for Rus- sian and Japanese. In addition, other functions have been implemented for several of these languages.

In this paper we discuss our philosophy, which constrained our design decisions; elaborate some specific functions a morphological component should support; survey some competing approaches; describe our technique, which provides the necessary functionality while meet- ing the other design constraints; and support our approach by characterizing our success in developing/testing processors for various combinations of language and function.

D E S I G N P H I L O S O P H Y

Before we set about designing our morphological processor, we elaborated our philosophical commitments regarding an NLP system. These include: (1) multilingual application, (2) fault-tolerant, fail-soft behavior, (3) rule reversibility, (4) disparate functionality, (5) inherent parallelism, (6) grammatical clarity,

and (7) rigorous testing.

M U L T I L I N G U A L A P P L I C A T I O N

The algorithms and their software instantiation must a d m i t application to any h u m a n language likely to confront our system; these include the languages of major demographic a n d / o r economic significance (and their relatives). Our selected representatives are English, French, German, Spanish, Russian, Arabic, Chinese, and Japanese.

F A U L T - T O L E R A N T , F A I L - S O F T B E H A V I O R

Real-world N L P applications, whether for text or interactive dialog, will be confronted with numerous errors of various types. As far as users are concerned, guaranteed failure in the presence of any error is intolerable: a system must overcome simple mistakes without discern- able problems. F o r example, insignificant typing a n d / o r spelling mistakes should be ignored, as should minor morphological blunders. Users do not like to be asked for corrections of (seemingly) simple mistakes, and of course printed texts cannot be queried in any case. In the presence of more serious problems, performance should degrade only gradually. This is nothing more than a c o m m i t m e n t to

understanding

the utterance, r a t h e r than

punishing

the user for errors in it. We contend t h a t human-like fault-tolerant, fail-soft behavior must be incorporated in the fundamental design of a system: it cannot be tacked-on after system development. Creating an applied system for a " p e r f e c t " natural language t h a t is hypothesized, but never observed, is misguided, aside from being wasteful.

R U L E R E V E R S I B I L I T Y

(2)

forms as input), w h i l e a synthesizer should never produce t h e m as o u t p u t . Since these two functions are, therefore, r a t h e r different processes, one m u s t search for a means to distinguish the rules (linguistic descriptions) from the strategies (linguistic processes, called g r a m m a r s ) controll- ing their application: the former can be reversible, while the latter m i g h t not be.

D I S P A R A T E F U N C T I O N A L I T Y

Analysis and synthesis constitute two obvious linguistic processes ( g r a m m a r s imposed upon rule sets). There are, however, even more processes t h a n these of interest in a practical setting: t h a t is, there m a y be a n u m b e r of gram- m a r s built from a single set of linguistic rules, as we d e m o n s t r a t e below. Thus, a processor m u s t a d m i t the simultaneous instantiation of a number of g r a m m a r s in order to be called general. I N H E R E N T P A R A L L E L I S M

Acceptable r u n t i m e performance in any significant real-world N L P setting is now understood to require i m p l e m e n t a t i o n on parallel machines. Thus, g r a m m a r s m u s t be inherently suited to parallel execution, and such oppor- tunities m u s t somehow be representable in the formalism in which the g r a m m a r s are expressed. G R A M M A T I C A L C L A R I T Y

In a n y real-world application, the n u m b e r of linguistic rules and the complexities of g r a m m a r s imposed upon t h e m is considerable. Successful i m p l e m e n t a t i o n and maintenance thus requires t h a t the g r a m m a r s be clearly and concisely stated, however powerful they m a y be. N o t only must the rule formalism be relatively clean and simple, but also a g r a m m a r m u s t be viewable at various levels of detail. Variable g r a n u l a r i t y in the presentation enhances the o p p o r t u n i t y for comprehensibility.

R I G O R O U S T E S T I N G

In order to become practical, a system must a d m i t - - and have undergone -- rigorous testing. It is not enough to develop a micro- implementation, then claim t h a t the system can be scaled up to become a real application (presumably to be tested on end-users). More than just a philosophical c o m m i t m e n t to the idea of testing, this requires t h a t the system ac- tually be fast enough

during the development

phase

t h a t thorough testing (of g r a m m a r s , etc.) can take place at t h a t time. If its speed is gla- cial during the development phase, a complex system cannot be completed, and its practicality will never be shown.

M O R P H O L O G I C A L P R O C E S S E S We have, so far, identified five interesting morphological processes: analysis, guessing, synthesis, defaulting, and coding. T h e first two concern comprehension; the other three, production.

A N A L Y S I S

Morphological analysis is a relatively well- understood notion -- which is not to say t h a t there is a g r e e m e n t concerning w h a t the result should be, or how it should be produced. But, generally speaking, analysis is agreed to involve the decomposition of a surface-form string (usually in English called a

word)

into its con- stituent base-form m o r p h e m e s and their func- tional a t t a c h m e n t s ; this m a y be finer-grained, as when each m o r p h e m e is associated with lexical or other linguistic information, and indeed the process is usually understood to imply access to a stored lexicon of morphemes, in order to cor- rectly identify those contained in the string. Analysis is assumed to p e r f o r m this decomposition according to a set of language-specific strategies (i.e., a g r a m m a r ) limiting the possible decompositions.

G U E S S I N G

In keeping with a c o m m i t m e n t to fault- tolerant, fail-soft behavior, a system must, e.g., deal with unknown words in an u t t t e r a n c e by making plausible guesses regarding their morphological, lexical, syntactic, and even semantic properties. A morphological g u e s s i n g g r a m m a r , p r e s u m a b l y operating after the analysis g r a m m a r has failed, m u s t embody heuristic strategies, and these m a y well differ from those of the analyzer, even though the rule stock upon which they draw might be identical. F o r example, while the analyzer must, sooner or later, entertain all possible decomposition hypotheses, the guesser might best be constrained to produce only the " m o s t l i k e l y / p l a u s i b l e " hypotheses.

S Y N T H E S I S

(3)

D E F A U L T I N G

Guessing is relevant to end-users, dealing as it does with unknown words in an input utterance. Developers, on the other hand, faced with teaching the system [how to synthesize only] the correct surface forms of words being defined, can make use of additional functions, such as a d e f a u l t i n g g r a m m a r . Given a root or stem form and p a r t of speech, a lexical defaulter can propose the most likely inflectional paradigm for a word. This is better than requir- ing the lexicographer to type in the variants, or manually encode the paradigm in some other fashion: greater h u m a n reliability is experienced when validating good guesses and correcting a few wrong ones than when entering everything from scratch.

C O D I N G

When the lexical defaulter guesses incor- rectly, a c o d i n g g r a m m a r could render all potential inflectional paradigms (as permitted by the language), from which the lexicographer could select the correct one(s) for the newly- defined word. This is desirable, because greater human reliability is experienced in selecting from among possible inflections than in producing all and only the correct variants.

S U R V E Y

One of the more popular frameworks for morphological processing is the two-level model developed by Koskenniemi [1983], modified and extended by K a r t t u n e n [1983]. Two-level models are fully reversible, performing both analysis and synthesis, and correspond to finite-state machines, hence they appear to enjoy some computational advantages. However, there are both theoretical and practical problems. It appears t h a t the model is not sufficiently powerful to account for some h u m a n languages (e.g., Icelandic, which exhibits recursive vowel harmony). The model can be decidedly inelegant in some respects (e.g., in handling alternations such as the English

nominate/nominee by positing a

"lexical e n t r y "

nomin).

There is a very substantial time penalty involved in compiling two- level g r a m m a r s (it may be measured in hours), which impedes rapid debugging and testing. Finally, the " a d v a n t a g e s " of reversibility are ar- guable, for reasons discussed above and below.

Affix-stripping models are commonly employed, especially for English [Slocum, 1981]. The IBM system [Byrd, 1983] also uses affix rules in a strictly word-based model of morphology, b u t the rules are unordered and thus only marginally may be said to constitute a gram- mar; certainly, morphosyntactic behavior is not highlighted. These two systems were developed for English analysis only; they are not reversible

in any sense, and have hot been tested on other languages. A serendipitous situation exists, in t h a t they have a certain degree of fault-tolerance built in, though this was not a goal of the design/implementation process.

In order to elucidate morphosyntax, some approaches use a simple (e.g., two-level, or similar) model to account for orthographic behavior, and a " g r a m m a r of w o r d s " to analyze m o r p h e m e se- quences syntactically (e.g., [Bear, 1986], [Russell et al., 1986]). It does not seem t h a t these approaches, as described, lend themselves to any sort of reversal, or other form of rule-sharing; furthermore, only analysis is mentioned as an application.

Even simpler models have employed allomorphic segmentation; m o r p h o s y n t a x may be ap- p r o x i m a t e d by a longest-match heuristic [Pounder and K o m m e n d a , 1986] or defined by a finite-state g r a m m a r [Bennett and Slocum, 1985]. The required entry of every allomorphic v a r i a n t of every word is b o t h a theoretical and a practical disadvantage, but r u n t i m e processing is speeded up as a positive consequence due to total neglect of spelling changes. F a u l t - t o l e r a n t behavior is not built-in, but can be almost trivially added (whereas, in m a n y other models, it would be difficult to incorporate). T h e systems cited here are used for analysis only.

No previous models of morphology seem to have been used for anything but analysis, and occasionally synthesis; the other three functions mentioned above, and others t h a t might exist, are neglected. Although the a u t h o r has discussed other functionality in earlier work [Slocum and Bennett, 1982], even there the morphological processors used for analysis, synthesis, and defaulting/coding were distinct, being implemented by entirely different software modules, and shared only the dictionary entries.

T H E N A B U T E C H N I Q U E

In Nabu, rules are separate from the strategies imposed upon them. Rules may be t h o u g h t of as declarative in nature; they are or- ganized in an inheritance hierarchy. The " g r a m m a r s of w o r d s " imposed upon them, however, may be t h o u g h t of as procedures -- ac- tually, dataflow networks.

R U L E H I E R A R C H Y

(4)

level division imposed by the lingusts is t h a t of category: the part-of-speech of the surface-forms for which the rule subset is relevant. F o r languages t h a t distinguish inflection from derivation, our linguists have generally found it con- venient to divide rules at the third level in t e r m s of these two classes. Other t h a n the top-level language division, however, the structure of the hierarchy is entirely at the discretion of the responsible linguists. Consequently, we have observed t h a t different linguists - each responsible for the morphological rules (and g r a m m a r s ) of entire languages - prefer and employ different organizational principles.

Rules m a y also be t h o u g h t of as declarative in nature -- though, in fact, for the purposes of maintenance they e m b o d y certain procedural be- haviors such as a self-test facility. A morphological rule is an equation between one letter- string+feature-set, which we call a g l o s s e m e , and another. One side of the equation describes w h a t might be called the " s u r f a c e " side of a glosseme, as it represents an encoding of information nearer to (but not necessarily at!) the surface-string level: all this really means is t h a t relatively more information is expressed in the string p a r t t h a n in the feature part. T h e other side, in turn, describes w h a t might be called the " b a s e " glosseme, as it represents an encoding of information closer to (but not necessarily at!) the base-form level, with relatively less information expressed in the string p a r t than in the feature part. It is i m p o r t a n t to note t h a t the in- formation content of the two sides is the same - only the form of the information changes. This is why we classify a rule as an e q u a t i o n , and this also admits rule reversibility in the implementation.

As a trivial example of such an equation, consider the English inflectional rule

[ " + s " 0] -~ [ " + " (NOUN PL)].

T h e " s u r f a c e " side, on the left, describes a glosseme string whose last character is the letter s [by means of the p a t t e r n "-t-s"] and places no constraints on the glosseme's feature set [by means of the e m p t y list 0]. The " b a s e " side, on the right, describes an equivalent glosseme whose string lacks the final letter 8 [by means of the p a t t e r n " + " ] and constrains the feature set [by means of the two features (NOUN PL)]. Ex- ecuted in one direction, this rule removes an 8 and adds the two features (NOUN PL); reversed, the rule removes the two features ( N O U N PL) and adds the m o r p h e m e s. Obviously, this English rule conveys the notion t h a t a N O U N m a y be inflected for P L u r a l by means of the [bound] m o r p h e m e s at its right end.

T h e plus sign ( + ) in a p a t t e r n is used to in- dicate whether prefixation or suffixation of the glosseme string is being described, depending on whether the p a t t e r n precedes or follows it; or a p a t t e r n m a y describe an entire glosseme string via omission of the sign. In a p a t t e r n , al- phabetic case is i m p o r t a n t : a lower-case letter signifies a constant, and m u s t be m a t c h e d by the same letter (in a n y case) in the glosseme; an upper-case letter signifies a restricted variable, and m u s t be m a t c h e d by some letter f r o m the set over which it is defined. Thus, for example, in English rules we use the letter V to stand for Vowel; C, for Consonant; and G, for G e m m i n a t - ing consonant. (Of course, the variable restrictions are entirely a r b i t r a r y as far as N a b u is concerned. Each linguist defines sets of variables and their m a t c h restrictions according to taste and the language at hand.)

If the s a m e variable a p p e a r s more t h a n once in a p a t t e r n , it is required to m a t c h the same letter in the glosseme: the equation

[ " + G G i n g " 01 = [ " + G " (VERB PRPL)]

thus describes doubling of the last consonant in an English verb, before suffixation of the present participial m o r p h e m e . A n o t h e r facility is required for convenience in describing alternation. In G e r m a n , for example, certain morphological operations involve the u m l a u t i n g of vowels; thus, an u n m a r k e d vowel on the " b a s e " side m a y require r e p l a c e m e n t by an u m l a u t e d vowel on the " s u r f a c e " side. If only one vowel behaved this way, this would be no problem: the corresponding letters would simply a p p e a r in their places in the patterns. B u t there are three vowels t h a t behave like this in G e r m a n (a, o, and u) in the identical context. In order to eliminate the need for tripling the size of the rule base otherwise required to describe this phenomenon, we provide the linguists with a means for pairing-up variables, so t h a t a character m a t c h i n g one m a y be replaced by the corresponding character in the o t h e r ' s set. M a n y other languages exhibit this kind of alternation, making this a useful technique.

A character in a p a t t e r n string matches one and only one character in a glosseme string. Generally speaking, the characters appearing in a p a t t e r n string are those of the language being described, as one would expect. Some languages, however, lack a case distinction -- Arabic, for example -- rendering the variable notation prob- lematic. In this situation, upper-case letters from the R o m a n a l p h a b e t are used to represent variables in rules.

(5)

ward. Rule execution is a m a t t e r of matching according to one side of the equation and (if t h a t is successful) transforming according to the other side. So long as every rule is truly an equation (and only a h u m a n linguist can decide this, ex- cept in trivial cases) then every rule is reversible -- t h a t is, can be used for comprehension as well as production, because thc interpreter can transform a rule's o u t p u t back into the original input. In Nabu, as noted curler, there are cur- rently two comprehension processes (analysis and guessing) and three production processes (synthesis, defaulting, and coding). But neither collections of rules nor their hierarchical structure describes such functionality.

C O N T R O L G R A P H S

A morphological g r a m m a r is an execution strategy imposed upon a bocly of morphological rules. Bearing in mind that morphological rules can apply to the outputs of other rules (consider

the derivational process in English, as for example w h e n the word derivational is constructed from derive + ation + al), and t h a t such compounding is not free, but linguistically constrained, it is obvious t h a t the compounding constraints -- as well as the individual rules themselves -- must be accounted for. In Nabu, an execution strategy is represented as a dataflow network, which we loosely term a control graph.

A control graph is a collection of nodes con- nected by directed arcs. Each node is composed of a bundle of morphological rules, which may be applied when input reaches t h a t node, and whose o u t p u t may be passed across arcs to other nodes. T h e r e will be one or more designated start n o d e s , where input appears and processing begins (conceptually, in parallel, if there are multiple s t a r t nodes). F r o m each s t a r t node, there will be a p a t h to one or more designated t e r m i n a l n o d e s , where processing ends and the graph's o u t p u t is emitted. The path may be of a r b i t r a r y length; s t a r t nodes may themselves be terminal nodes. In analysis, encountering a terminal node entails dictionary look-up; in synthesis, o u t p u t of a surface-form.

T h e r e are two types of arcs, S u c c e s s and F a i l u r e ; the former are arcs across which successfully-applied rules will pass their output, and the latter are arcs across which the original input to a node is passed, should n o rule inside it be successful. A successful rule is one whose input side ( " s u r f a c e " or " b a s e , " depending on whether the graph is engaged in comprehension or production) matches the input glosseme, and whose o u t p u t side represents the same information as was present in the input, only refor- mulated.

Conceptually, the rule~ inside a node may be

executed in s e r i e s or in p a r a l l e l . The linguist controls this by setting a flag in each node; thus, some nodes may fire rules in parallel, while others fire their rules serially. (In serial nodes, rule execution terminates as soon as one rule succeeds; there is no backup.) In either case, all success arcs are traversed in parallel, or else all failure arcs are traversed in parallel, depending on whether any rule(s) succeeded -- meaning all possible morphological analyses are produced for later consideration.

T o take a simple example, consider the word derivations, and assume t h a t neither it nor the singular form derivation is in the dictionary. An English analyzer graph might have multiple s t a r t nodes, one of which is intended (by the linguist) for inflected nouns. T h e input glosseme

[ " d e r i v a t i o n s " O]

is thus passed to the node P L U R A L - N O U N , which contains, among others, the rule

[ " + s " 0l = [ " + " (NOUN PL)].

T h e suffix p a t t e r n " + s " matches the glosseme string, and no features are required to be present in the glosseme, thus this rule succeeds and produces the o u t p u t glosseme

[ " d e r i v a t i o n " (NOUN PL)[.

If P L U R A L - N O U N has been m a r k e d as a terminal node (in addition to being a s t a r t node), then dictionary look-up will take place. If so, by our assumption above it fails. Our hypothetical graph contains a Success arc from PLURAL- NOUN to D E V E R B A L - N O U N , which contains, among others, the rule

["+ation" (NOUN)I = ("+e" (VERB (DERIVE NOUN +ION))).

When (and if) this rule fires, it would match the suffix (ation) in the glosseme string derivation, and the feature (NOUN), and therefore transform t h a t glosseme into

pL)f'.'derive" (VERB (DERIVE NOUN + I O N )

(6)

ner. If D E V E R B A L - N O U N has been marked as a terminal node, then dictionary look-up will take place: the V E R B entry

derive

is retrieved. The glosseme feature li~t, in addition to indicat- ing the main-entry ( " d e r i v e " ) and lexical category (VERB) to look for in the dictionary, contains sufficient information to allow transformation of the stored entry for

derive

into a representation of its surface-form realization

(derivations),

in terms of syntactic category (NOUN), sub- categorization features (PL), and semantics (the meaning of

derive,

transformed by +ION).

T h e r e remains only one problem to account for: t h a t of compositional vs. non-compositional readings. W o r d s with strictly non-compositional readings (e.g.,

fruitful,

which is not computable from

fruit

and

ful)

simply must be stored in the dictionary; this is not a contentious claim. Words with strictly compositional readings (e.g.,

derivation,

which is computable from

derive

and

ation)

may or may not be stored in the dictionary: this is an efficiency consideration, based on the usual time vs. space trade-off, and is also not subject to significant debate -- at least, not with respect to theoretical implications.

The problem arises for cases where both compositional and non-compositional readings exist for the same word (e.g.,

largely).

In such situa- tions, the non-compositional reading must of course be stored, b u t it would be nice if this did not

require

storage of the compositional reading as well. In Nabu, we solve this by means of a D E C O M P O S A B L E flag t h a t must appear within the non-compositional definition of a word, in case that word also has a compositional reading which is to be computed. During the course of morphological analysis, performing successful dictionary look-up (at a " t e r m i n a l " node) will affect subsequent processing: if the DECOM- P O S A B L E flag is noted, then the glosseme just submitted to dictionary .look-up will

also

be passed across any success arcs leaving t h a t node, for further analysis. In the absence of a D E C O M P O S A B L E flag, successful dictionary look-up marks a truly terminal case, and the results are returned (possibly along with readings from other paths in the graph) forthwith.

The operations of other types of control graphs are analogous to those of the analyzer. The principles are the same, with small excep- tions (some noted above), and so are not ex- emplified here.

P R O C E S S O R S I M P L E M E N T E D In addition to the rule and graph interpreters per se, delivered to MCC shareholders in mid-1985, Nabu includes a variety of tools sup- porting the development, testing, maintenance, and multilingual documentation of morphological rule hierarchies and grammars. These tools

(not described in this paper, for reasons of space) have been used to create a great m a n y rules and several morphological processors. Non-terminals included, the English morph-rule hierarchy numbers 626 nodes; French, 434; German, 493; Spanish, 1395; and Arabic, 882. In addition to these m a t u r e rule hierarchies, some preliminary work has been done on the Russian and Japanese rule hierarchies.

A N A L Y Z E R S

T h e English analyzer is complete with respect to inflection; it has been successfully tested on, a m o n g other things, the entire collection of inflectional variants presented in Webster's Seventh N e w Collegiate Dictionary (ca. 42,500 nouns, 8,750 verbs, and 13,250 adjectives). It also accounts for the great bulk of English derivation, as determined by various word frequency lists, and is undergoing gradual, evolutionary extension to the missing (low- frequency) affixes and their combinations. A first version of this g r a m m a r was delivered to M C C shareholders in mid-1985, followed by upgrades in 1986 and 1987. The current analyzer numbers 20 nodes and 60 arcs.

As mentioned earlier, a complete Arabic morphological analyzer exists; so far as we are aware, it accounts for all morphological phenomena in the language -- no mean feat, for a language in which a single root form could in theory be realized as over 200,000 surface forms, and in which morphemes are frequently discon- tinuous (i.e., cannot be described by simple af- fixation models) [Aristar, 1987]. This 371-node, l133-arc analyzer was delivered to MCC shareholders in mid-1986, and may represent the first complete analyzer ever produced for Arabic. The French and G e r m a n analyzers are complete with respect to inflection (highly irregular forms, like

sein

in German, naturally excepted). The former numbers 71 nodes and 121 arcs; the latter, 54 nodes and 79 arcs. The 19-node, 17- arc Spanish analyzer is nearly complete with respect to inflection; adjectives remain a tem- p o r a r y exception. With respect to verbs, for example, it has been tested on an extensive list of conjugated verbs [Noble and Lacasa, 1980], com- prising over 6,000 surface forms, and in the first such test it was 970"/0 accurate.

G U E S S E R S

(7)

like Arabic, are morphologically very rich, ad- mitting as a consequence very strong predictions.

D E F A U L T E R

A 21-node, 23-arc English defaulting graph exists. It seems to be complete (insofar as such a processor might be), in that it constitutes a seemingly adequate component of a dictionary entry coding tool.

C O N C L U S I O N S

Morphological grammars in Nabu are able to account for all compositional readings of arbitrarily-complex surface-forms in a wide range of languages. Furthermot'e, the formalism and development environment are reasonably comfortable. These claims are supported by our implementation and large-scale testing of several diverse grammars.

For philosophical reasons, we are opposed to the idea that grammars (as opposed to individual rules) must be reversible: even if it were not for the need of five-fold rather than merely dual functionality, the need for fault-tolerance in a practical system, without consequent fault- exhibition, argues for separate analysis and synthesis grammars. We also point out that, in our implementations, the [non-reversible] control graphs tend to be much smaller in size than the hierarchies of [reversible] rules, hence the storage penalty for "redundancy" is inconsequential.

R E F E R E N C E S

Aristar, A., "Unification and the Computa- tional Analysis of Arabic," Computers and Translation 2, 2, (April-June) 1987, pp. 67-75.

Bear, J., "A Morphological Recognizer with Syntactic and Phonological Rules," Proceedings of COLING86, Bonn, 1986, pp. 272-276.

Bennett, W.S., and J. Sloeum, "The LRC Machine Translation System," Computational Linguistics 11 (2-3), 1985, pp. 111-121.

Byrd, R.J., "Word Formation in Natural Language Processing Systems," Proceedings of the 8th IJCAI, Karlsruhe, 1983, pp. 704-706.

Karttunen, L., "Kimmo - A General Mor- phological Processor," Texas Linguistic Forum 22, 1983, pp. 165-186.

Koskenniemi, K., "Two-level Model for Mor- phological Analysis," Proceedings of the 8th IJ- CAI, Karlsruhe, 1983, pp. 683-685.

Noble, J., and J. Lacasa. Handbook of Spanish Verbs. Iowa State University Press, Ames, Iowa, 1980.

Pounder, A., and M. Kommenda, "Morphological Analysis for a German Text-to- Speech System," Proceedings of COLING86, Bonn, 1986, pp. 263-268.

Russell, G.J., S.G. Pulman, G.D. Ritchie, and A.W. Black, "A Dictionary and Morphologi- cal Analyser for English," Proceedings of COLING86, Bonn, 1986, pp. 277-279.

Slocum, J., "An English Affix Analyzer with Intermediate Dictionary Look-up," Working Paper LRC-81-1, Linguistics Research Center, University of Texas, February 1981.